Predictive Database Performance Optimization Using Machine Learning-Driven Query Workload Modeling
Year: 2021
Type: article
Abstract: The database systems of modern time have to be running on very dynamic and heterogeneous workloads where the old rule-based and cost-based optimization methods may tend not to sustain steady performance. In this paper, a predictive database performance optimization framework can be proposed that is implemented on the basis of machine learning-based query workload modeling. The method makes use of historical records of execution, structure characteristics of queries, and model characteristics of runtime resources to create forecasting models that can predict query latency and system resource consumption. The framework incorporates feature extraction, workload classification, temporal workload pattern learning, and the supervised machine learning models to predict the occurrence of performance bottlenecks. The system obtains complex nonlinear relations among query properties and run-time performance to support proactive optimization techniques such as adaptive indexing, memory set up, and execution plan optimization. The hybrid modeling methods that require a combination of plan-level and operator-level modeling can further improve the accuracy of prediction and the generalization of various workloads. Benchmark workloads can be used to experimentally evaluate the improvements in prediction accuracy and query execution performance relative to traditional cost based optimizers. Findings have shown a decrease in query latency, better throughput and a better efficiency in resources during dynamic workload situations. The framework proposed will lead to the creation of self-optimizing smart databases systems, which will be able to continuously adapt to cloud and distributed systems.
Language: English
Subfield: Information Systems
Field: Computer Science
Domain: Physical Sciences
Open Access status: bronze
OpenAlex ID: W7132844025